DETOUR (Text): leaderboard
Metric: Accuracy (%) on the 254 DETOUR queries without an input file, mean of three runs of the full harness (DETOUR tip-of-the-tongue known-item search: the evaluated primary agent must name the entity behind an underspecified, sometimes deliberately erroneous human recollection by querying a fixed Gemini-2.5-Pro memory agent (which answers only from a hidden memory file) and a web search tool; GPT-4o judges the final answer against the annotated entity; temperature 0; each model searches with a DuckDuckGo tool except GPT-5, which used OpenAI's native web-search-preview tool); higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 66.85 | #91 |
| 2 | Claude Sonnet 4.5 | 59.04 | #138 |
| 3 | Kimi K2 (Thinking) | 45.62 | #236 (Kimi K2) |
| 4 | Gemini 2.5 Pro | 42.13 | #145 |
| 5 | Llama 4 Maverick | 23.11 | #451 |
| 6 | Qwen 3 Next 80B A3B (Thinking) | 20.41 | #306 (Qwen 3 Next 80B A3B) |
| 7 | Llama 4 Scout | 16.69 | #646 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=detour-text · How It Works · Data refreshed daily, snapshot 2026-10-11.